Research Of Feature Extraction Method Based On Gist Feature And Manifold Learning

In the researching field of pattern recognition and computer vision, the effective andreasonable extraction of the image feature and maping high dimensional sample feature to lowdimensional subspace play key role to the classification and recognition step. This dissertation,deeply studied graph based manifold learning dimensionality reduction algorithm, human facefeature extraction method, and gist feature extraction method, improved methods are proposedfor them respectivly. The author proposed new face recognition method and buildingrecognition method based on these improved dimensionality reduction algorithm, human facefeature extraction method, and gist feature extraction method.The main research contents and research achievement of this dissertation are as follows:1. Neighbor graph construction plays a vital role on the performance of graph basedmanifold learning dimensionality reduction algorithm. The author proposed a new neighborgraph construction method named Corresponding Columns Based Graph (CCG) constructionmethod to overcome the defect which the traditional k-nearest neighbors Graph constructionmethod didn’t take the2D structure of sample matrix into account when constructing theneighbor graph for dimensionality reduction algorithms. Different from the k-nearest neighborsGraph construction method transform each sample into their vectorial form and determine k-nearest neighbors for each of vectorized sample, the CCG compute the k-nearest columnneighbors of each column of each sample matrix. Then, the author brought a sample threshold,if the number of column neighbors between two samples exceed the sample threshold, the twosamples become neighbors. The author applied CCG to Locality Preserving Projection (LPP)and Marginal Fisher Analysis (MFA). Experimental results show that the LPP and the MFAwhich using the CCG achieved much better performance than original algorithms, and anattractive property of CCG is columns’noise immunity.2. The author proposed Corresponding Block Based Graph (CBG) base on the CCG. TheCBG overcome the shortcoming of the traditional k-nearest neighbors Graph constructionmethod fertherly, and the CBG is robust to non-uniform illumination in some externt. The CBGdevided a sample matrix into several blocks, and compute k corresponding neighbor blocks for each blocks. The CBG determine whether two samples can become neighbors according to thenumber of neighbor blocks between them. The author applied CBG to Locality PreservingProjection (LPP), and proposed CBG-LPP. The author carried out a series of experiments onthe ORL, the YALE, and the CMU PIE face databases, and CBG-LPP achieved better resultsthan LPP.3. We need to predefine neighbor parameter k for all samples before computing theneighbor graph. This may cause parameter selection difficult, and set the same parameter k forall samples is Inappropriate. To overcome the parameter selection difficult, and take the2Dstructure of sample matrix into account when constructing the neighbor graph, the authorproposed Samples’InnerStructureBasedGraph (SISG) construction method. The SISG dependon the newly proposed sample similarities to determine the neighbors of each sample, and thesample similarities are calculated from the samples’ inner structure information. The SISG is ageneral neighbor graph construction method, and it can be applied to many graph basedmanifold learning dimensionality reduction algorithm. The author applied the SISG to LPP, andproposed SISG-LPP. The author conducted a series of experiments on several well-known facedatabases, and Experimental results show that SISG-LPP outperformed LPP, CCG-LPP, andCBG-LPP.4. The author proposed a newly defined Histograms of Oriented Gradients (HOG) calledLocality Sensitive Histograms of Oriented Gradients (LSHOG). The LSHOG overcome thedefect which face feature extracted by the classical HOG can’t reflect the2D structure of faceimages, and also reduce the influence of non-uniform illumination and occlusion on the faceimages. The LSHOG build a histogram of gradient orientations at each pixel location, whichtakes the contributions of gradient orientations of every image pixel into account. But, thecontribution to the histogram of gradient orientations at a pixel location of all image pixel isnot same. The author add a locality sensitive parameter to each occurrence of a gradientdirection value. The locality sensitive parameter can make the farther the distance from the thepixel location where the histogram at, the smaller contribution of a pixel makes. The LSHOGwas utilized to extract feature vectros of face images from AR, CMU PIE, ORL, YALE facedatabases. Then, several manifold learning algorithm were utilized to reduce the dimensionalityof face feature vectros. Experimental results prove that the LSHOG based face recognitionmethod can achieves better recognition accuracy than the HOG based face recognition method.5. Siagian and Itti’s gist feature extraction model was proposed for scene recognition task.Li and Allinson applied this gist feature extraction model to building recognition task. But,building recognition task is different from scene recognition task, because buildings may encounter the non-uniform illumination. The author proposed subregion’s multiscale gistfeature (MS-gist) extraction model. Siagian and Itti’s gist feature extraction model computed34sub-channel feature maps from the whole building image, and divided feature maps into sub-regions. The gist vector of the building image is obtained by take mean of each sub-regions.The MS-gist divided the building image into×blocks at first, and compute feature mapfrom each block. The gist vector extracted by MS-gist is obtained by take mean of each featuremap of each block.6. The author combined MS-gist with manifold learning algorithm based on Samples’Inner Structure Based Graph, and proposed a new building recognition method called BuildingRecognition on Subregion’s Multiscale Gist Feature Extraction and Samples’ Inner StructureBased Dimensionality Reduction. The author’s building recognition method is a two stagemodel. In the first stage, MS-gist was utilize to extract gist feature vectors of building images.In the second stage, the author use SISG-LPP to reduce the dimensionality of gist featurevectors. To evaluate the building recognition method, the author conduct several experimentson Sheffield building databases, and achieved satisfactory results.